Application of Machine Learning-Based Models in Early Detection of Production Anomalies in Oil and Gas Wells

The early detection of production anomalies in petroleum wells is critical for optimizing operational efficiency, reducing downtime, and mitigating environmental risks, particularly in complex and prolific regions like the Niger Delta. This study investigates the application of both unsupervised and supervised artificial intelligence models for anomaly detection using historical well production data. A comprehensive exploratory data analysis of 15,634 records revealed significant variability and skewness in key parameters such as downhole pressure and oil production, justifying the need for advanced detection techniques. Several machine learning models were evaluated, including Isolation Forest, One-Class Support Vector Machine (SVM), and Copula-Based Outlier Detection (COPOD) as unsupervised methods, together with Random Forest and XGBoost as supervised methods. The unsupervised models demonstrated varying degrees of success; One-Class SVM achieved the highest recall (0.879), effectively capturing most anomalies but with lower precision, while COPOD offered a more balanced performance (F1-score: 0.66). The supervised models initially achieved perfect classification metrics (F1-score: 1.0); this was traced to a data-leakage error in which the target column was inadvertently retained among the input features. After removing the leaked feature and re-evaluating with time-series cross-validation, XGBoost achieved the best supervised performance (F1-score: 0.85, AUC: 0.95) and Random Forest a precision of 0.95, giving realistic rather than perfect results. Reconstruction-based methods such as autoencoders were also explored and are identified as a promising direction for future evaluation. The results underscore a fundamental trade-off: unsupervised models provide scalable, label-free solutions ideal for initial screening across diverse assets, whereas supervised models deliver superior accuracy when high-quality labeled data is available. The study concludes that a hybrid framework, leveraging unsupervised methods for broad monitoring and supervised models for confirmed anomaly types, offers the most pragmatic and effective strategy for deployment in the dynamic operational environment of the Niger Delta. This approach promises to enhance proactive well management, support predictive maintenance, and contribute to more sustainable and economically efficient petroleum production.